Hyperspectral Image Classification Based On Few-Shot Learning

The classification of ground objects is the key technology and necessary steps for a variety of applications of remote sensing observation to earth.Hyperspectral imaging technology effectively combines spectroscopy and imaging technology,so that hyperspectral image contain high-resolution spectral information as well as large coverage of spatial information,and become important data source for object classification tasks.The popular hyperspectral image classification method is based on the deep network model with excellent learning ability,and the accuracy that meets the actual application can be obtained when each type of training sample is sufficient.However,due to the particularity of remote sensing tasks,ground surveys are time-consuming and labor-intensive,which makes it difficult for the quantity and quality of labeled sample data to meet the requirements of deep network training,and restricts the popularization and development of ground object classification tasks based on hyperspectral images.Aiming at the problem of training samples,this paper further studies the hyperspectral image classification model based on deep networks,and proposes a study of hyperspectral image classification methods based on few-sample learning.The main work of this paper is as follows:First,in order to solve the problem of low overall classification accuracy caused by insufficient number of labeled samples,i.e.the problem of few-shot learning,a hyperspectral image classification method based on a two-stage relationship learning network is proposed.The network uses the convolution-based feature embedding module to learn the spatial and spectral features of the image,and obtains the relationship score between the known sample and the unknown sample through the relationship learning module,thereby inferring the category of the unknown sample.The relational learning module allows the network to learn generalization category knowledge in other hyperspectral data sets for information compensation,and learn personalized knowledge in the hyperspectral image to be processed for fine training,therefore,the classification accuracy of hyperspectral image under small samples can be improved through information additional compensation and information depth mining.Second,in order to solve the problem of low accuracy of some categories caused by the imbalance in the number of labeled samples between different categories,i.e.the problem of large categories eating small categories,a hyperspectral image classification method based on a specific task learning network is proposed.The specific task learning module in the algorithm can generate a set of parameters based on each input sample,and specifically adjust the feature extraction network,so that it can better match each type of data.Through the specific task learning module,useful information can be learned from different categories equally,and unbiased feature extraction can be realized,thereby improving the classification accuracy of hyperspectral image in the case of imbalanced training samples between different categories.In summary,this article has conducted an in-depth analysis of the existing problems of the insufficient number of training samples and the imbalance between different categories in the current hyperspectral image classification task,and proposed related solutions.Combining the relevant knowledge of deep learning,a hyperspectral image classification method based on few-sample learning is proposed,and the effectiveness of the proposed method is verified by related experiments using multiple open hyperspectral image data.